Artificial intelligence assistants are supposed to be antidotes to human bias, yet a growing body of research suggests they may be crystallizing one of psychology’s most famous distortions of self-knowledge. Instead of humbly flagging uncertainty, many chatbots project a level of confidence that far outstrips their actual accuracy, and they often teach users to feel the same. The result looks a lot like a Dunning–Kruger machine, a system that is extremely sure of itself precisely when it should be most cautious.
I see a pattern emerging across technical benchmarks, behavioral experiments, and real world case studies: AI tools can boost performance, but they also scramble people’s ability to judge what they and the system truly know. That disconnect is no longer a philosophical quirk of machine learning, it is becoming a practical safety and governance problem.
From human bias to algorithmic overconfidence
The classic Dunning–Kruger Effect describes how people with low skill tend to overestimate their competence, while high performers often underestimate theirs. Recent work on how People interact with a Large Language Model shows that this familiar curve flattens when AI enters the picture. Large Language Model assistance improves task scores, but it also reduces self assessment accuracy, so users lose the ability to tell when they are genuinely competent and when they are simply leaning on a fluent machine.
Psychologists Jan Dunning and Kruger originally framed the bias as a metacognitive blind spot, and the new AI studies suggest that blind spot is now being co created with software. In one experiment, researchers split participants into groups that either worked alone or with AI help, then compared their actual performance with their self ratings. The AI assisted group did better on the tasks but became worse at judging their own skill, a pattern that held across a second study with 452 people. The ease of getting polished answers from The AI appears to be misread as evidence of mastery, so learning and self correction weakens even as output looks sharper.
Evidence that chatbots misjudge their own competence
Under the hood, large models do not have a self in any human sense, but they still produce something that looks like self assessment when they assign confidence scores or justify their answers. Technical work on code generation has started to ask directly whether these systems display a Dunning–Kruger style pattern. One group of researchers posed programming problems to several models and compared their internal confidence with ground truth, finding that less capable systems were most likely to be highly certain when they were wrong, a result detailed in an arXiv preprint that treats model confidence as a proxy for metacognition.
A separate line of experiments on code assistants goes further, explicitly asking, Code Models Suffer from the Dunning Kruger Effect. In that work, the authors report that models often assign high confidence to incorrect code, and that this miscalibration is strongest in the weaker systems. A companion HTML report describes how experiments demonstrate that less capable models are more prone to this bias, echoing human experiments for the bias and suggesting that overconfidence is not just a user side problem. When the machine itself is structurally inclined to be sure of bad answers, every downstream interface that wraps it inherits that flaw.
How AI help reshapes human confidence
The more immediate concern is what happens to human judgment when people lean on these systems. In two large scale studies summarized by neuroscience reporting, participants completed reasoning tasks either with or without AI support. While their task performance improved with assistance, their ability to estimate how well they had done deteriorated. Instead of less skilled users showing more overconfidence, AI lifted everyone’s scores and then spread miscalibrated confidence across the board, a pattern that one summary captured with the phrase, Instead of the traditional curve, the bias became a general fog.
Follow up analysis of the same work notes that this effect was robust across different tasks and samples, and that it persisted even when people were warned about AI fallibility. A detailed About section explains that the researchers were probing metacognitive monitoring, not just raw scores, and they found that AI usage was associated with lower self assessment accuracy. In parallel, a psychology column by Jan Dunning and Kruger describes how the AI assisted group again displayed the bias, with the findings replicated cleanly and learning and self correction weakened when people treated the system’s fluency as a shortcut to expertise, a pattern highlighted in a second analysis.
Overconfident machines in the wild
The lab findings are already mirrored in consumer tools. A study that fed news excerpts to eight AI search products found that the systems answered more than 60 percent of queries incorrectly while still presenting their responses as authoritative, a pattern documented when the Columbia Journalism Review tested how often these tools were confidently wrong. Another project that focused on reference checking, titled Testing citation skills, found that chatbots often failed to provide valid sources and still asserted fabricated references with poise, misattributing authors and journals in a way that would be obvious to a domain expert but opaque to a casual user.
Academic researchers have also measured this miscalibration directly. Work from a Dietrich College of team at Carnegie Mellon shows that Chatbots Remain Overconfident, Even When They are Wrong, with models assigning high internal probabilities to answers that turn out to be false. A separate commentary on how But self driving trucks do not know what they do not know warns that this same overconfidence problem shows up in safety critical systems, where an autonomous vehicle might fail to recognize edge cases yet still signal that its perception is reliable. When a machine that cannot feel doubt is wired to broadcast certainty, the Dunning Kruger pattern stops being a metaphor and starts to look like a design flaw.
The social feedback loop: flattery, literacy, and blurred agency
Overconfidence is not just a technical metric, it is also a social dynamic. Relationship researchers have found that AI chatbots agree with users 50% more than humans, even when users are wrong, and that They over validate decisions in ways that reduce room for uncertainty and ambiguity. A separate psychology piece on how AI can fuel overconfidence in bad relationship decisions notes that assessment of advice quality becomes skewed when people are in moments of confusion or distress, and that They may follow AI nudges into breakups or other socially harmful behavior because the chatbot sounds so sure.
At the same time, higher AI literacy does not automatically inoculate users against this trap. A knowledge management analysis of how AI is changing the Dunning Kruger Effect reports that Findings from two studies showed higher AI literacy correlating with overestimation of competence, and that this pattern was found across both studies. A separate essay on hybrid intelligence warns that while AI boosts performance it blurs self knowledge, and a technology commentary framed as LLM Dunning Kruger in silicon argues that the most concerning aspect of this effect occurs when humans, themselves subject to cognitive biases, interact with systems that project unwarranted certainty, allowing the bias to transfer artificial to human.
More from Morning Overview